slimdex-mcp

A local MCP server that helps coding agents retrieve code narrowly instead of reading whole files into context. An agent asks Slimdex for a specific outline, line range, symbol body, or reference list, rather than loading a file to find one thing.

Status: pre-1.0. It works on the repos it has been run against, but it has not been published, packaged, or validated broadly. Read What's actually verified before relying on it.

The retrieval guidance below also ships in the server's MCP instructions, so clients inject it into the model's context automatically.

Recommended agent flow

brief first, at the very start of a session — one call that reports what the repo is, where recent sessions were digging, and which saved conclusions still match the code (stale ones flagged), so a fresh chat starts informed instead of blank. Then get_context("Foo") to answer "what is this, who calls it, what does it depend on" in one response. To understand a whole area rather than one symbol, context_pack("how does auth work") runs the entire exploration server-side and hands back a single bounded bundle — the relevant symbols, how they connect, and the top bodies — so you spend one call and one transcript entry instead of ten. Don't know the name, only what it does? — search_intent("parse the config file") ranks symbols by intent with BM25, no embeddings. Drop to get_symbol_context for one body (it flags itself if the file drifted from the index, so you don't re-read to check), get_file_skeleton for a file's shape, and read_lines when you need exact source. Before editing a symbol, find_tests on it to see what covers it; to rewrite a whole function, replace_symbol (you send only the new body — the old code isn't re-sent just to locate the edit). Use batch to bundle several lookups. Every search tool takes limit (default 20) and offset.

Response budgeting: get_context sections are opt-in via include (default: definition, signature, callers, imports — add body or dependents explicitly), callers are capped by callerLimit, and the response is bounded by maxChars (default 12,000). Every cap that trips prints an explicit notice (showing 3 of 68, truncated at maxChars=...) rather than dropping data silently. get_symbol_context caps its span with maxLines the same way, and memory_list returns the newest 50 facts unless told otherwise, as ~150-char previews rather than whole bodies (memory_get ids:[...] expands them, full:true dumps everything). On an 18-fact store that is the difference between ~4,100 and ~18,600 chars in the call every session opens with.

Config: <root>/.slimdex.json (optional)

{
  "ignoreDirs": ["fixtures", "backend/src/main/resources/static/assets"],
  "extensions": [".astro", ".vue"],
  "suffixes": [".stories.mdx"],
  "exclude": ["generated/", "legacy/vendor"],
  "maxFileBytes": 2000000
}

suffixes matches a filename ending, for file types an extension can't identify. Salesforce metadata sidecars ship as a built-in: AccountSvc.cls-meta.xml, panel.js-meta.xml and Account.object-meta.xml are indexed, while pom.xml, web.xml and manifest/package.xml are not — adding .xml to extensions would have pulled in every config tree in the repo. Suffix-matched files are indexed for search and read reach, not symbols.

Merged on top of the built-in ignore list (node_modules, dist, .venv, .svelte-kit, Pods, .pytest_cache, …). An ignoreDirs entry is either a bare name, matching any directory so called at any depth, or a path containing /, anchored at the repo root and respecting directory boundaries (src/gen will not also ignore src/generated). index_repo echoes what it loaded and warns about unknown keys, wrong types, or invalid JSON, so a typo'd config isn't silently indistinguishable from none.

Build output usually needs no config at all. Beyond the directory list, any file whose lines run past ~5,000 characters is treated as minified build output and left out of the index — bundlers strip newlines, and hand-written source doesn't look like that. This catches what a name list structurally cannot: a hash-named bundle (index-B7xK2p9q.js) inside a directory called assets. assets, public and static are deliberately not ignored by name, because real source lives in them; index_repo reports the count as skipped(minified build output): N.

How the token saving works

There's no compression trick. The saving is behavioral: these tools let an agent retrieve outlines, ranges, and locations instead of whole files, and the persistent index means repeat lookups hit a cached query rather than a re-read.

Two later sessions, run by different models on different repo shapes, added real-world numbers to the original report:

Multi-file web app, bug-fix session (GPT-5.3-Codex). 19 credits reported with slimdex; the model's own estimate for the same scope without it: 45–70 credits. Math: 19/45 → 19/70 ≈ 58–73% cheaper. The counterfactual is the model's estimate, not a measured A/B — directional.

Single giant file (folio-app: one 6,200-line, 313 KB app.js). Slimdex's own stats: ~34,000 chars across 8 calls ≈ 9–10k tokens — one skeleton (213 signatures), then bodies of only ~12 relevant functions, 9 of them fetched in a single get_symbol_context names:[...] call. The naive path: 313 KB ≈ 78–85k tokens across 3–4 forced full reads. Math: ~10k vs ~80k ≈ ~70k tokens saved, an 85–90% reduction on exploration. The bug's diagnosis (an export path with no matching import path) was visible from the skeleton's signatures before a single body was opened.

Together they sketch the scaling law: the saving scales with how much irrelevant code the naive path would drag in. One giant file is the best case; a normal repo lands around half to two-thirds cheaper; a repo of tiny files breaks even. Same standing caveats as everything here: stats count chars, not tokens (÷3.5–4), and single sessions are evidence, not benchmarks.

Both figures above measure reading only, which is the cheaper half. Output costs roughly 4–5× input, so an undisciplined edit wastes more than an undisciplined read: rewriting a whole function through a generic edit tool means re-sending the entire old body purely so the tool can locate it. replace_symbol addresses by name and that cost disappears. stats reports this alongside follow-through, because the leak is otherwise invisible — the expensive path still produces a correct edit, so nothing signals that you overpaid:

write discipline:
  replace_symbol: 0 call(s), 0 symbol(s) rewritten by name
  changed outside slimdex: 12 file(s)
  pre-edit checks (find_tests/dep_graph/get_context/changed_files): 0

External edits are inferred from content hashes moving between two index_repo runs, so the number is honest about its limits: it sees that bytes changed, never which tool changed them, and a human editing in another window counts too.

The realistic whole-workflow band

The figures above are single-scenario exploration numbers — the best case, where the naive path would have dragged in the most irrelevant code. Averaged across a whole real workday, not just the exploration slice, the band settles lower:

• ~55–60% on navigation-heavy work — reading and understanding a codebase, where narrow retrieval replaces whole-file reads most often.
• ~45% on output-heavy work — churning out new code, where more of the cost is generation the server doesn't touch (though replace_symbol now shaves the write side too).
• ~50% averaged over regular day-to-day use. The saving compounds the more sessions run through it, because brief and memory mean each new chat starts informed instead of re-deriving the repo from zero.

Use it regularly across sessions in your IDE for the best of this.

Treat these as one data point, not a benchmark. Single repo, single task, one A/B run each, self-measured, no repetitions or variance. Your mileage depends heavily on whether your agent actually reaches for the narrow tools instead of falling back to reading files — which varies by client and model. The method is repeatable if you want to check it: run the same task in two fresh sessions, one instructed to use only Slimdex and one instructed to avoid it, and compare /status cache-write.

What's actually verified

Being explicit, since the rest of this README is easy to over-read.

Covered by the unit suite (npm test runs 224 tests across 23 files):

• Symbol extraction across JS/TS (incl. class and object-literal methods), Python, Go, Rust, Java/C#, and comment skipping — symbols.test.ts

• Import extraction for JS import/require/export-from, Python, Rust

• Block extraction, brace-scoped and indentation-scoped, with string/comment awareness (quotes, templates, //, /* */, full-line #) — extractBlock.test.ts

• Import resolution, external-module classification, reverse-edge dependents, Mermaid emission, and root-BFS depth scoping — graph.test.ts

• Search match format, pagination without overlap, per-line occurrence counting, exact totals, regex escaping/rejection — search.test.ts

• Opaque cursor round-tripping and malformed-cursor rejection; parser-backend fallback — pagination.test.ts

• Outline declaration detection vs. control flow — outline.test.ts

• get_symbol_context maxLines budgeting and truncation notice

• String/comment masking and brace-depth tracking — lexer.test.ts

• Per-language extraction for all twelve supported languages — languages.test.ts

• The index cache returns the same object until the index is rewritten

• .slimdex.json loading: every key applied through a real index build, plus the failure modes (invalid JSON, unknown keys, wrong types) each producing a visible warning instead of silence — config.test.ts

• changed_files against a real temporary git repository: hunk→symbol attribution, untracked files, explicit base refs, and formatting; skips cleanly when git isn't installed — git.test.ts

• The file watcher, with real fs events: a save is debounced, reindexed, and lands in the on-disk index — watch.test.ts

• Graph edges beyond imports: name-reference edges for import-less code (class→used-class, interface→implementation via dependents, trigger→handler) and declarative-wiring edges from repo XML (metadata-binding→class), with comment/string mentions excluded and per-build caching — apexgraph.test.ts

• The in-memory file cache serves repeats without re-reading and always serves fresh content after an on-disk change — fscache.test.ts

• Test-file detection across JS/TS/Python/Go/Ruby/Java/C# conventions, with Windows separators normalized and ordinary source (latest.ts, Contest.java) not misflagged — testlink.test.ts

• The write side: replacing a symbol's block, trailing code preserved, and CRLF vs LF line endings kept so an edit isn't reflowed into a whole-file diff — edit.test.ts

• Memory staleness: a fact is marked live when it names a symbol/file that still exists, flagged stale only when every code mention is gone, and left unflagged for prose — plus brief composition — brief.test.ts

• Intent search: camelCase/snake_case tokenization, and BM25 ranking that surfaces a differently-named symbol by its intent words while scoring an unrelated query to nothing — intent.test.ts

• Freshness: a file newer than its indexed mtime reads as stale (line numbers may be off), a matching mtime reads as fresh, and a missing file never cries stale — freshness.test.ts

• context_pack assembly: header + ranked symbols + bodies in one bundle, the no-match message, char-budget gating that still guarantees the first body, and the symbols-limit cap — pack.test.ts

• The architecture digest: covered files modified after the digest read as stale, a newer digest reads clean, coverage-scope and directory-prefix filtering, and the rendered fresh/stale verdict — digest.test.ts

Covered end to end, through the real MCP server (integration.test.ts spawns the server over stdio against a temporary fixture repo and asserts on output): index_repo, repo_map, read_lines, get_file_skeleton, outline_file, get_symbol_context, find_definition, find_references, find_tests (the hit and the no-coverage warning), search_intent (intent ranking), context_pack (one-call bundle), digest_save/digest_get (round trip with freshness verdict), get_context (including its maxChars cap), dep_graph (imports + mermaid), batch, search_code, search_symbols, stats, brief, replace_symbol (write-then-query round trip and the unknown-symbol refusal), the memory_save/search/list/delete round trip, the path-escape guard, and the not-found paths.

CI runs the build and both suites on Ubuntu + Windows, Node 20 and 22.

Caveat on the watcher test: recursive fs.watch is platform-dependent, so watch.test.ts degrades to a logged skip on filesystems that never deliver an event — same behavior as the watcher itself. On Windows, macOS, and current Linux it asserts the full save→reindex path.

npm run smoke still exists but proves only that the pipeline is alive — the correctness assertions live in integration.test.ts.

Verified by inspection: src/ contains no network calls — no code leaves your machine. This one you can check yourself: grep -rE "fetch\(|https?://|axios|http\.request" src/.

Longer documentation

In docs/:

• tool-guide.md — every tool explained twice (technically and in plain words) with an example each, the combined workflow, and how mtime-based persistence works
• tool-guide.html — the same guide as a styled, self-contained page for the browser
• token-savings-report.md — the original A/B measurement, its method, and how to repeat it
• agent-brain.md — the full operating discipline as a readable document
• agent-brain-slim.md — the one to drop into a repo as CLAUDE.md / AGENTS.md. Self-contained and one page: savings ladder, question→tool table, memory discipline, session hygiene, honest limits, env knobs. Same coverage as the full document at ~30% of the prose, because the tool rules are dense tables rather than paragraphs the server already injects.

Language coverage

Two measurements, because fixtures alone prove very little.

Fixtures — one per language, counting the declarations a developer would actually navigate to: 65/65 found, 0 false positives, pinned by test/languages.test.ts.

Real third-party code — extraction run over ~11,800 files from several hundred real packages (React, Babel, Remix, Socket.io, Playwright, Three.js, Emotion, zod, ajv …) and compared against an independently written heuristic for what counts as a declaration: 95.9% recall. Reproduce it yourself:

npm run audit -- ./node_modules            # or any directory of code you didn't write

That number is a floor, not a grade — the truth heuristic counts some non-declarations, so real recall is a little higher. What it's for is catching regressions and finding the next real gap.

About frameworks

Almost nothing that failed the audit was framework-specific. Frameworks add annotations, decorators and conventions; they rarely invent syntax. Handle the language and the frameworks come with it — fflib's Application/Domain/Selector/ Service/UnitOfWork layers extract completely (129 declarations) without a single fflib-aware rule.

The one genuine exception is test DSLs. A vitest/jest/mocha/RSpec file often has no top-level declarations at all, so entire test directories used to index to nothing. describe/it/test titles are now indexed as kind test, which is what you actually navigate to in a test file.

Framework semantics are recovered wherever the reference exists somewhere in the repo, through two extra edge sources in the graph:

• Name-reference edges, for languages that have no import statement (e.g. Apex): if one file's code — comments and strings masked out — mentions a top-level type defined in another file, that's an edge. This is what makes implements answerable as "who implements this interface", and links a trigger to the handler class it news up.
• Declarative-wiring edges: bindings that frameworks keep in configuration rather than code (custom-metadata records, flow definitions) usually live in the repo as XML with the type name as an element value. Repo XML is scanned for known type names — XML comments excluded — and each hit becomes a metadata-file → class edge, so dependents answers "what wires this up".

Both scans are cached per index build and cost nothing on repos without such files. Pinned by apexgraph.test.ts. What no static reader can see is a binding that exists only in a live system — configured in a running org or database and never retrieved into the repo. If it's not in the repo in any form, there is no edge to draw; search the type name instead.

Performance

Cold index is a full parse; warm is an mtime check per file. Measured on Windows, Node 24.

The index is held in memory and invalidated by the index file's mtime. Without that cache every tool call re-read and re-parsed the whole index — about 20 ms of dead weight per call on the 5,000-file repo, and it grew with the repo.

find_references is the slowest tool at scale because it is a textual scan, not an index lookup — but a literal pre-filter now skips the line-split and per-line regex for any file whose raw source doesn't contain the searched name, which on a typical repo is most of them. Scope with pathPrefix to cut the remaining file reads when you know roughly where to look.

File contents are also served from a byte-bounded in-memory LRU (64 MB, validated by mtime+size per hit), so the second scan of a repo — and the skeleton→read_lines→context sequence agents actually perform on one file — costs a stat() instead of a read.

Memory across sessions

memory_save writes to <root>/.slimdex/memory.json, which outlives the process — a fact saved in one chat is readable in the next, by a different client, after a restart. Chat and editor share one store only when both point at the same SLIMDEX_ROOT.

Nothing is captured automatically: the server never sees your conversation, so the agent has to decide what's worth keeping. The shipped instructions tell it to read memory first in a new session and to save decisions, constraints and gotchas as it learns them — but that's guidance to the model, not a guarantee.

Known limitations

• Symbol extraction is regex-based and heuristic, not a parser or LSP. It can miss unusual declarations, and find_references is a textual match that may include same-named but unrelated identifiers.
• Symbol and outline extraction now run against a masked copy of each line, with string and comment contents blanked out, so declaration-shaped prose inside a template literal is no longer indexed as code. Declarations are also depth-aware: a const x = () => … or type X = … counts only at top level, because locals inside a function body are not things anyone navigates to. Class methods are still indexed at their nesting depth.
• An inline Python # comment containing a brace can still confuse block extraction (# is also the JS private-field sigil, so it can't be stripped blindly).
• changed_files attributes a hunk to the nearest preceding declaration — right for a normal function body, approximate for code between declarations. Treat it as blast radius, not a call graph.
• search_code reports an exact total but stops at an internal scan cap on very large result sets, printing N+ (scan cap reached) rather than a confident wrong number.
• Language support is uneven: JS/TS is the best-covered. C-family and Ruby, formerly the thinnest, gained dedicated rules (free functions, Foo::bar definitions, function-like macros, attr_*); the remaining soft spots are advanced C++ shapes — templates split across lines, operator overloads.
• For LSP-grade precision you'd swap the parser for tree-sitter or a language server. src/parser.ts is the seam: a Parser interface selected by SLIMDEX_PARSER, with the regex parser as the only implementation that ships. A tree-sitter backend would drop in there without touching any tool or the index format. It is not built — per-language grammars trade away the "installs instantly, runs offline, zero config" property.

Deliberately not built

Ideas evaluated and rejected, with reasoning — these are design opinions, not measured results:

• Symbol-ID dictionaries (S42 → path) — MCP has no client-side expansion layer, so the model receives an opaque token it must spend another call to resolve.
• Token-budget managers / cost estimators — chars/4 estimates are unreliable across tokenizers, and auto-compressing on a bad estimate can drop data the model needed.
• Delta / "already-sent, see response #5" caching — after context compaction the earlier payload is gone, so the reference resolves to nothing.
• Embeddings / semantic search — large dependency footprint; possible future optional flag, not a default.
• A tree-sitter parser backend — this is the one that would close the remaining ~4%, and it was costed rather than hand-waved: web-tree-sitter is WASM so it needs no native compilation, but the grammars (tree-sitter-wasms) are 51.7 MB unpacked against ~4.5 MB for the whole current install. Evaluated and declined at 95.9% measured recall, because "installs in a second, runs offline, no configuration" is the property this server exists to have. src/parser.ts remains the seam if that calculus ever changes — a backend drops in there without touching a tool or the index format.

Install

Not published to npm. There is no npx slimdex-mcp. Build from source:

git clone https://github.com/Siddhukaushik/slimdex-mcp
cd slimdex-mcp
npm install
npm run build      # produces dist/index.js
npm test           # vitest unit suite

Verify it runs end to end against a repo:

npm run smoke                                # this repo
node smoke-test.mjs "C:/path/to/some/repo"   # any other

Environment variables

The persistent cache

Per repository, Slimdex writes to <repo>/.slimdex/:

• index.json — the code index (mtime-invalidated per file, and discarded wholesale when the index format version changes, so a stale index built by an older extractor is never reused)
• memory.json — saved memory facts
• stats.json — per-tool usage counters

The directory ignores itself: a * .gitignore is written inside it (the node_modules/.cache trick), so it never shows up in git status and you don't have to touch the repo's own .gitignore. Delete that inner file if you want to commit the cache.

Wiring it into MCP clients

MCP is a shared standard, so the same server should plug into any MCP-capable client. The project root is passed via SLIMDEX_ROOT (or as the first CLI arg).

Only Claude Code and Claude Desktop have actually been run. The others below are the standard config shape for each client, written from their documented format — they are untested here and may need adjustment.

Replace <ABS_PATH> with your build output, e.g. C:\path\to\slimdex-mcp\dist\index.js, and <REPO> with the repo to index.

No tuning required. The savings that matter are on by default in every client: memory facts list as previews, responses are terse, an identical re-read of an unchanged file answers with a pointer instead of the body, and several symbol edits go in one call. The env vars below are for opting out, or for lean — which trades a further ~8,700 chars/turn against routing a third of the tools through batch, so it is deliberately not the default.

Claude Code (CLI) — tested

claude mcp add slimdex --env SLIMDEX_ROOT=<REPO> -- node <ABS_PATH>

Claude Desktop — tested

%APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "slimdex": {
      "command": "node",
      "args": ["<ABS_PATH>"],
      "env": { "SLIMDEX_ROOT": "<REPO>" }
    }
  }
}

Codex CLI — tested

~/.codex/config.toml

[mcp_servers.slimdex]
command = 'C:\Program Files\nodejs\node.exe'
args = ['<ABS_PATH>']
startup_timeout_sec = 30

Registered globally like this, slimdex attaches to every Codex task and uses that task's working directory as the repo root — no SLIMDEX_ROOT needed. Codex launches the server with a restricted environment, so give command an absolute path to node rather than relying on PATH.

Cursor — untested

.cursor/mcp.json (project) or ~/.cursor/mcp.json (global)

{
  "mcpServers": {
    "slimdex": {
      "command": "node",
      "args": ["<ABS_PATH>"],
      "env": { "SLIMDEX_ROOT": "${workspaceFolder}" }
    }
  }
}

Windsurf — tested

~/.codeium/windsurf/mcp_config.json — same mcpServers shape as Cursor.

VS Code (Copilot / MCP) — tested

.vscode/mcp.json

{
  "servers": {
    "slimdex": {
      "command": "node",
      "args": ["<ABS_PATH>"],
      "env": { "SLIMDEX_ROOT": "${workspaceFolder}" }
    }
  }
}

Cline (VS Code extension) — tested

Cline settings → MCP Servers → add:

{
  "slimdex": {
    "command": "node",
    "args": ["<ABS_PATH>"],
    "env": { "SLIMDEX_ROOT": "<REPO>" }
  }
}

Zed — tested

settings.json → context_servers

{
  "context_servers": {
    "slimdex": {
      "command": { "path": "node", "args": ["<ABS_PATH>"], "env": { "SLIMDEX_ROOT": "<REPO>" } }
    }
  }
}

For clients that expose the workspace folder (Cursor, VS Code), ${workspaceFolder} keeps Slimdex pointed at the repo you have open.

Typical agent workflow

1. index_repo once at the start (faster on subsequent runs), then brief to pick up where past sessions left off with stale notes already flagged.
2. repo_map → get the lay of the land.
3. outline_file on a file of interest → pick line ranges.
4. read_lines for just those ranges.
5. find_definition / find_references / dep_graph to navigate.
6. find_tests before editing a symbol; replace_symbol to rewrite one without re-sending its old body.
7. memory_save decisions and gotchas so the next session starts informed.

License

MIT © 2026 Kael VK Inc. (Business Number 751569161 RC0001) — see LICENSE.

Provided as is, with no warranty and no support. If it doesn't build, doesn't run, or doesn't work on your setup, that's yours to carry — see the disclaimer in the license.